Convolutional Neural Network based Traffic-Sign Classifier Optimized for Edge Inference
B B Shabarinath, Muralidhar Pullakandam · 2020
Traffic-Sign Classification is a major task in self-driving cars as well as modern driving assisting systems can be deployed as an inference engine on the Field Programmable Gate Array (FPGA) combined with host processor-based edge device like Zynq Platform considering the property of dynamic reconfigurability of FPGA adoptable to architectural innovation. Hence this paper proposes an optimized Convolutional Neural Network (CNN) architecture based on VGGNet combined with image-preprocessing techniques. The proposed methodology utilizes pruning combined with post-training quantization-based optimization and obtains an accuracy loss of less than 1%. The architecture is trained, tested and validated using German Traffic Sign Detection Benchmark (GTSDB) with Google's TensorFlow framework obtaining a validation accuracy of 99.2% and test accuracy of 100% for novel input inference. The experimental results show the reduction in memory footprint of CNN model readily implementable on FPGA.